Original Source Here
Infrastructure and Data Foundation.
Trustworthy and secure AI models require a trustworthy and secure data foundation (as the saying goes, “garbage in, garbage out”). In considering the NAIRR implementation roadmap, we urge the Task Force to prioritize specific data infrastructure components that are most essential and practical to ensuring high quality and reliable AI outputs: intuitive data integration and pipeline transformations, audit logs, granular access controls, the ability to version and branch data, and collaboration features for annotating datasets and identifying addressable issues over time (e.g., statistical and other forms of unwanted data bias). With these key investments, the NAIRR effort has the potential to support, enrich, and grow the AI research community by providing a shared data environment, not just static datasets. The NAIRR data resource can enable tangible, trustworthy results through collaboration tools that allow for secure modifications, annotations, and improvements to datasets and the ability to share knowledge and performance metrics. These investments will enable the NAIRR infrastructure to scale and support cross-organization, cross-discipline research to springboard the U.S.’ AI capabilities.
Technical, Governance, and Cultural Awareness in Responsible AI.
Not only should the NAIRR incorporate best practices around AI Privacy and Civil Liberties (PCL), Bias, and Ethics, but it can — and should — enforce these best practices (e.g., stringent security and access controls) through a combination of a) technical infrastructure that facilitates enforcement of key data protection and responsible AI principles; b) governance that incentivizes and rewards best practices; and c) cultural awareness and discipline-specific frameworks to contextualize AI research and guide determinations of when and how AI applications can best align with the interests of impacted communities. Through our decades-long work developing and implementing data integration and management platforms built on granular security and privacy-preserving capabilities, we have demonstrated that well-engineered data infrastructure must provide the technical implementation measures that attach to and reinforce critical institutional governance measures. The challenges that AI research will ultimately address are not exclusively technological in nature, rather, they are techno-social and require an understanding of the cultural contexts, innovative data science practices, and institutional controls.
Problem Prioritization through Resource Allocation.
The Task Force is well-positioned to provide opinionated oversight that steers the NAIRR towards pressing, relevant problems and historically overlooked or underfunded initiatives. Similar to how the NSF approves grants, the Task Force can provide incentives and allocate resources (compute, data, etc.) based on problem evaluation. For example, the Task Force could identify AI safety research as a topic requiring more focused research and development. When industry alone researches this topic, it is likely to direct its attention to narrow, commercially-focused use cases. The NAIRR, in contrast, could expand the scope and insights produced by the AI research community. With the appropriate infrastructure and governance in place, the NAIRR has the potential to shed light on how AI resources are being used, discover whether the most important problems (as seen through a broader societal impact lens) are being researched, drive investment, and compound knowledge from across academic, government, and industry participants.
Recommendations to the NAIRR Task Force
Building off these three themes, we shared with the NAIRR Task Force several recommendations to proactively craft processes and policies that will sustain the NAIRR over the long-term. Prioritizing data quality and policies that incentivize and promote flexibility and a constantly improving technical infrastructure will empower the NAIRR with a sustainable program to harness the power of the U.S. technology sector to benefit the AI research community. We have summarized our key recommendations in the chart below:
Read our full response to the National Artificial Intelligence Research Resource Task Force RFI here.
Trending AI/ML Article Identified & Digested via Granola by Ramsey Elbasheer; a Machine-Driven RSS Bot